Are FPGAs Suitable for Edge Computing?

Saman Biookaghazadeh, Fengbo Ren, Ming Zhao

Abstract

The rapid growth of Internet-of-things (IoT) and artificial intelligence applications have called forth a new computing paradigm--edge computing. In this paper, we study the suitability of deploying FPGAs for edge computing from the perspectives of throughput sensitivity to workload size, architectural adaptiveness to algorithm characteristics, and energy efficiency. This goal is accomplished by conducting comparison experiments on an Intel Arria 10 GX1150 FPGA and an Nvidia Tesla K40m GPU. The experiment results imply that the key advantages of adopting FPGAs for edge computing over GPUs are three-fold: 1) FPGAs can provide a consistent throughput invariant to the size of application workload, which is critical to aggregating individual service requests from various IoT sensors; (2) FPGAs offer both spatial and temporal parallelism at a fine granularity and a massive scale, which guarantees a consistently high performance for accelerating both high-concurrency and high-dependency algorithms; and (3) FPGAs feature 3-4 times lower power consumption and up to 30.7 times better energy efficiency, offering better thermal stability and lower energy cost per functionality.

Introduction

As the Internet of Things (IoT) scales toward billions of devices, the traditional cloud-centric model faces significant challenges. Centralizing data processing creates latency and bandwidth bottlenecks that are unacceptable for time-sensitive applications. Edge computing addresses this by moving processing power closer to data sources. However, many edge servers currently rely on GPUs, which are optimized for high-throughput batch processing. This paper investigates whether Field-Programmable Gate Arrays (FPGAs) are better suited for the small, streaming, and unpredictable workloads typical of the edge.

Methodology

The study compares two high-end accelerators: an Intel Arria 10 GX1150 (FPGA) and an Nvidia Tesla K40m (GPU). Both devices were programmed using OpenCL to ensure a fair comparison. Evaluation metrics included throughput sensitivity to batch size, architectural adaptiveness to data and conditional dependencies, and energy efficiency (performance per watt). Benchmarks included multi-stage matrix multiplication ($A \times B \times C$) and custom kernels with varying loop dependencies.

Throughput Sensitivity to Workload Size

A critical requirement for edge computing is the ability to handle small, real-time data streams without the latency penalty of batching. GPUs typically require large batch sizes to hide communication overhead and saturate their parallel cores. In contrast, FPGAs utilize a pipelined architecture that processes data as it arrives.

Throughput vs. Batch Size for FPGA and GPU.

As shown in the experimental results, the FPGA maintains a consistent throughput regardless of the batch size. For batch sizes under 128, the FPGA significantly outperforms the GPU, demonstrating its superior suitability for low-latency edge workloads where waiting to accumulate large batches is not feasible.

Architectural Adaptiveness and Resilience

Edge algorithms often involve complex logic, including data dependencies and conditional branching. While GPUs excel at spatial parallelism (SIMD), they struggle with temporal dependencies and "branch divergence," where threads must wait or replay instructions. FPGAs provide both spatial and temporal parallelism at a fine granularity, allowing for the creation of dedicated hardware pipelines that handle dependencies efficiently.

Raw and Normalized Throughput under varying dependency levels.

When the degree of data dependency increased from 16 to 256 operations, the GPU's raw throughput plummeted by approximately 14x. Conversely, the FPGA's performance remained stable. In high-dependency scenarios, the FPGA achieved a normalized throughput (GFLOPS/fclk) significantly higher than the GPU, highlighting its architectural resilience to algorithmic complexity.

Energy Efficiency

Energy consumption is a primary constraint for edge deployments. The experiments revealed that the FPGA consumes 3–4 times less power than the GPU. In terms of energy efficiency (performance per watt), the FPGA was up to 30.7 times more efficient for small workloads and 11.8 times more efficient for high-dependency tasks. This efficiency translates to better thermal stability and lower operational costs.

Conclusion

FPGAs offer distinct advantages for edge computing, including predictable throughput for streaming data, resilience to complex algorithmic dependencies, and superior energy efficiency. While the GPU ecosystem currently offers greater ease of use, the performance and efficiency gains of FPGAs make them a compelling choice for the next generation of edge infrastructure. Future work will focus on expanding benchmarks to real-world edge algorithms and optimizing memory access patterns.

Summary updated

Abstract

The rapid growth of Internet-of-things (IoT) and artificial intelligence applications have called forth a new computing paradigm--edge computing. In this paper, we study the suitability of deploying FPGAs for edge computing from the perspectives of throughput sensitivity to workload size, architectural adaptiveness to algorithm characteristics, and energy efficiency. This goal is accomplished by conducting comparison experiments on an Intel Arria 10 GX1150 FPGA and an Nvidia Tesla K40m GPU. The experiment results imply that the key advantages of adopting FPGAs for edge computing over GPUs are three-fold: 1) FPGAs can provide a consistent throughput invariant to the size of application workload, which is critical to aggregating individual service requests from various IoT sensors; (2) FPGAs offer both spatial and temporal parallelism at a fine granularity and a massive scale, which guarantees a consistently high performance for accelerating both high-concurrency and high-dependency algorithms; and (3) FPGAs feature 3-4 times lower power consumption and up to 30.7 times better energy efficiency, offering better thermal stability and lower energy cost per functionality.
BibTeX
@misc{biookaghazadeh2018are,
  author = {Saman Biookaghazadeh and Fengbo Ren and Ming Zhao},
  title = {{Are FPGAs Suitable for Edge Computing?}},
  year = {2018},
  eprint = {1804.06404},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1804.06404},
}
Figure